liveBook · Manning

Skill library

As explained in Designing AI Agents

A collection of Skills maintained for selection and reuse by an agent system.

A skill library makes reusable procedures available for agent selection. Keeping it reliable requires tracking which Skills were eligible, what was selected, and how the run performed—not just counting uses. section 7.4.5

Maintain the library as an evidence-backed lifecycle:

  • Record task conditions, eligible and selected versions, agent and Harness versions, dependencies, outcomes, costs, failures, and corrections. section 7.4.5
  • Review for duplicates, uncovered tasks, misrouting, conflicts, dependency changes, context cost, and outdated versions. Choose to keep, refine, merge, demote, deprecate, or retire. section 7.4.5
  • Treat low use as a reason to review, not automatically demote; critical but rare Skills may still be needed. section 7.4.5
  • Make withdrawal evidence-backed and reversible, with a rollback version; require authorized approval for changes affecting safety or production boundaries. section 7.4.5
Listing 7.14 Operating evidence and reversible withdrawal
@dataclass(frozen=True)
class SkillRunEvidence:
    task_family: str
    legal_versions: tuple[str, ...]
    selected_version: str | None
    agent_version: str
    harness_version: str
    dependencies: dict[str, str]
    outcome: str
    evidence_ids: tuple[str, ...]
 
 
@dataclass(frozen=True)
class WithdrawalEvidence:
    reason: str
    evidence_ids: tuple[str, ...]
    reversible: bool
    rollback_version: str
    approval_id: str | None = None
 
 
def withdraw_skill(
    registry,
    skill,
    evidence,
) -> bool:
    if not evidence.reason or not evidence.evidence_ids:
        return False
    if not registry.valid_evidence(evidence.evidence_ids, skill):
        return False
    if not evidence.reversible: #A
        return False
    if not registry.has_version(
        skill.skill_id,
        evidence.rollback_version,
    ):
        return False
    protected_scope = registry.is_protected(skill) #B
    if protected_scope and not registry.valid_approval(
        evidence.approval_id,
        skill,
    ):
        return False
    with registry.atomic_change(): #C
        registry.archive(skill, evidence.rollback_version)
        registry.remove_from_routes(skill.skill_id, skill.version)
        registry.record_withdrawal(skill, evidence)
        skill.state = SkillState.DEPRECATED
    return True

A registry-facing withdrawal example checks evidence and reversibility before changing Skill availability. section 7.4.5

How Designing AI Agents covers it

How deep it goes
Details operating records, review actions, and reversible withdrawal, but does not specify a complete admission procedure.
Examples
Uses run records, periodic portfolio reviews, and evidence-checked withdrawal with a rollback version.
What you can do after
Operate a Skill portfolio using run evidence, review criteria, and reversible withdrawal.
How it connects
Connects Skill routing with evaluation, dependency health, versioning, and approval controls.

This explainer was generated by Manning's knowledge graph from the text of Designing AI Agents. The section links point to the parts of the book it draws on.